Open19 to challenge OCP in creating a hardware standard for data centres

by

21 September 2016
Yuval Bachar of LinkedIn

From 16 servers to 100,000 – that is the range of environments that LinkedIn is aiming to cover with Open19, a new data centre standards initiative that approaches the idea of “open” differently from the Open Compute Project (OCP).

Announced in July 2016, Open19 defines an architecture and a standard form factor for compute, storage and networking in any 19-inch rack environment.

Making a case for Open19 at the DatacentreDynamics Converged SE Asia event in September, Yuval Bachar, principal engineer in the global infrastructure and strategy team for LinkedIn, said the goal is to build data centres better by optimising base rack costs, speeding up rack integration and creating a framework for innovation.

The basic building block for the platform is a 19-inch rack. “We don’t define the rack; you can use any rack. It only needs to be 19-inch, which is the most common and prevalent technology,” said Bachar.

The severs can be in any of three form factors starting with half U – the brick, double wide brick or the double high brick. The configuration, said Bachar, is designed to fit every motherboard in the world and eliminate lock-in by any supplier.

“The open design allows for innovation to be pushed to the supplier, as long as they work with the shared form factor,” said Bachar. Companies can retain their intellectual property as they are not required to share their designs with the community. This, he believes, will help drive innovation in the Open19 ecosystem.

OCP, on the other hand, comes from a different point of view – that of the mega data centre operators, said Bachar. And it works for them because their environment justifies it.

With the OCP approach, servers are pre-designed by data centre operators to fit their applications and manufactured by single or multiple ODMs (original design manufacturers).

“Under OCP, different companies contribute hardware and software designs to the shared community, which has resulted in different specialty rack configurations and a variety of form factors. You have a 21-inch rack that Facebook contributed, a power base that Microsoft contributed, but none of it is standard.”

And, he pointed out, while the designs are shared with the community, not everybody can use them because they may not operate on the same scale as the mega data centre operators.

Another area of differentiation between Open19 and OPC is disaggregation.

“With OCP, you have rack-level cooling, power and controllers, and you don’t replicate that environment then you have a problem,” said Bachar. “With Open19, each element in the architecture is self-sustained. There is no centralised hardware-based solution to hold the rack together; no controller or centralised management on the rack itself.”

For example, each server can be managed, run and rebooted independently. Each element is also able to cool itself although to cater to a diversified environment, some implementations may feature cooling doors and self-cooling racks.

The only thing that is centralised is power distribution and this is done for efficiency,” said Bachar. There is no separate power supply for each server. Instead, there is a power shelf with 12V snap-on power cables for power distribution.

Open19 blind-mate snap-on connectivity is one aspect of the architecture that LinkedIn spent considerable effort in developing, said Bachar.

The snap-on power cables deliver 20-250 watts per server as a baseline, while blind mate 100GBbps data cables provide four bidirectional channels of up to 25GBbps each. The system can be assembled to support different configurations, much like assembling Lego bricks, said Bachar. For example, snap-on cables can deliver 200Gbps for every double wide brick (1U) and 400Gbps for a double high brick.

The connectivity creates an environment which allows for fast integration of the rack, said Bachar. Except for a few screws needed to install components such as the power shelf, most of the integration process is tool-less, thanks to the use of snap-on cables.

Today, it takes about six hours for a trained technician to do the integration, whether on-site or off-site, said Bachar. “We want to reduce it to one hour and a half. That is the way to scale from a 10,000-server environment to a 20,000-server environment and beyond.”